Re: [kbuild-all] [linux-next:master 6373/10489] fs/xfs/xfs_log_cil.c:897:1-10: second lock on line 900 (fwd)

["Darrick J. Wong" <djwong@kernel.org>]

On Thu, Jun 17, 2021 at 07:59:00PM +0100, Matthew Wilcox wrote:
> On Thu, Jun 17, 2021 at 11:50:44AM -0700, Darrick J. Wong wrote:
> > On Thu, Jun 17, 2021 at 08:28:24PM +0200, Julia Lawall wrote:
> > > cocci warnings: (new ones prefixed by >>)
> > > >> fs/xfs/xfs_log_cil.c:897:1-10: second lock on line 900
> > > 
> > > 5fd9256ce156ef Dave Chinner      2021-06-03  891  	/*
> > > 5fd9256ce156ef Dave Chinner      2021-06-03  892  	 * If the checkpoint spans multiple iclogs, wait for all previous
> > > cb1acb3f324636 Dave Chinner      2021-06-04  893  	 * iclogs to complete before we submit the commit_iclog. In this case,
> > > cb1acb3f324636 Dave Chinner      2021-06-04  894  	 * the commit_iclog write needs to issue a pre-flush so that the
> > > cb1acb3f324636 Dave Chinner      2021-06-04  895  	 * ordering is correctly preserved down to stable storage.
> > > 5fd9256ce156ef Dave Chinner      2021-06-03  896  	 */
> > > 5fd9256ce156ef Dave Chinner      2021-06-03 @897  	spin_lock(&log->l_icloglock);
> > > cb1acb3f324636 Dave Chinner      2021-06-04  898  	if (ctx->start_lsn != commit_lsn) {
> > > 5fd9256ce156ef Dave Chinner      2021-06-03  899  		xlog_wait_on_iclog(commit_iclog->ic_prev);
> > > cb1acb3f324636 Dave Chinner      2021-06-04 @900  		spin_lock(&log->l_icloglock);
> > 
> > xlog_wait_on_commit drops l_icloglock, either directly or via xlog_wait.
> > It looks odd (perhaps there should be a comment?) but at least in theory
> > the functions are annotated so I guess that means the static checking
> > doesn't know that commit_iclog->ic_log == log?
> 
> I think it's hard for a tool to reach into fs/xfs/xfs_log.c and look for
> the __releases annotation on the definition of xlog_wait_on_commit().
> Should we also annotate the prototype in fs/xfs/xfs_log_priv.h ?
> 
> For example,
> 
> void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
>                                  struct inode *inode)
>         __releases(&inode->i_lock);

That depends on whether or not amending the declaration in that manner
actually satisfies the checking tool?  Ah, I see, __releases is a macro
that only expands to anything if __CHECKER__, which is probably why the
actual checker tool doesn't see this, and possibly why gcc can't
complain about the mismatch between declaration and definition.

--D